Inspiration
Like a lot of students, we often struggle to get the right intuition for maths problems. We end up memorising the method, or not getting it at all, and that comes back to bite us when we realise we have no idea what we're actually doing.
Take $y'' + 4y = \cos 2x$ and $y'' + 4y = \cos 3x$. Most of us can grind out both solutions, but far fewer could tell you before solving which one grows without limit, and why. That's the gap we wanted to close.
What it does
Picture it AI helps students understand the intuition behind the questions on their problem sheets, so the sheets teach more than a method to copy.
A lecturer uploads a problem sheet. For each question, the student first predicts something about the answer from a picture built only from what the question gives them. For example: which solution grows, which way a paddle wheel turns in a field, or whether an area increases as a slice moves up.
If they're wrong, they don't get the answer. They get a consequence they can check for themselves, and they try again. Once the prediction is right, they solve the question on paper and type their answer, which is checked in code. Only then does the full picture appear.
How we built it
We spent a couple of hours going over different ideas, going to the workshops and talking to people from Reply. Then we spent all night and the morning turning the idea into a real product, building it with Claude.
We split into three tracks (engine, web page and AI agents) that worked in parallel against one shared set of data shapes:
- Engine: Python and FastAPI, with SymPy and NumPy. It checks predictions and answers and computes every claim a picture makes (for example that a flux comes to exactly x) before any student sees it.
- Agents: a team of Claude agents. One reads the PDF, one decides which questions are worth a picture, one designs the check, one reviews it as a teacher would, and others give hints and follow-up questions. The agents that talk to students never see the answers.
- Page: plain HTML and JavaScript, with 2D and 3D pictures.
The lecturer approves every question before students see it.
Challenges we ran into
- Coming up with something genuinely new. Plenty of tools already explain maths. We had to find something that makes students think rather than read.
- Stopping the agents from giving away the answer. Models love to be helpful. We added code checks that block any hint containing the answer or a correct feature.
- Incorrectly scaled diagrams. Early pictures were squashed or cut off, so we made the axes fit what is actually drawn.
- Pointless diagrams. Not every question needs a picture. An agent now decides whether a visual will actually help, a separate reviewer agent rejects ones that don't, and questions without a useful picture skip straight to the problem.
Accomplishments that we're proud of
- It never gives the answer away. We tested it with scripted "students" who begged for the answer, tried prompt injection and named formulas to trick it. Across 100 replies, none leaked the answer.
- The maths is checked by code, not by AI. Every picture's claim is computed before a student sees it, across 16 kinds of check, from which solution grows to the flux through a 3D surface.
- It goes beyond graphs. It covers curves, vector fields, regions and 3D surfaces, with pictures students can rotate and sliders they can drag.
What we learned
We learned a lot about agentic workflows: getting several specialised agents to hand work to each other, check each other, and know when to stop.
We also learned that many AI "tutors" give a false impression of learning. If the AI explains everything, it feels like you've understood, but you haven't done the thinking. So we deliberately didn't go down that path. Our AI never hands over the answer, and whether a student is right is always decided by code, not by a model.
What's next for Picture it AI
- Testing it with real students on real problem sheets, and seeing whether their intuition actually improves.
- Lecturer accounts, so uploads and class insights are properly protected.
- Making the pipeline reliable on more kinds of sheets and subjects, so a lecturer can upload almost any sheet and get useful checks back.
